OpenAI launches GPT-Live for voice conversations

OpenAI has launched GPT-Live, a voice model that can listen and speak at the same time, making conversations with ChatGPT feel more natural. It is now available globally on iOS, Android and ChatGPT.com.

The main difference lies in how it works. Instead of waiting for you to finish speaking before preparing a response, GPT-Live processes what you say while generating its own reply. It can pause, let you finish, respond with an “mhmm” or “I see,” and continue the conversation without turning it into a rigid sequence of turns.

A conversation that does not stop

Earlier voice systems had clear limitations. The first ones converted your voice into text, searched for an answer and then turned it back into audio. That process could lose information and add delay.

More recent models, such as Advanced Voice Mode, reduced latency by working directly with audio, but they still waited for you to stop speaking. A brief pause or background noise could make ChatGPT think you had finished and interrupt you.

GPT-Live uses a simultaneous two-way architecture, known as full-duplex. This means it listens and speaks continuously, making decisions several times per second about whether to respond, keep listening, pause, interrupt or use a tool.

In practice, you can say “wait a second” while you think, ask it to speak more slowly or interrupt it to correct a question. It can also keep a conversation going while conducting a search or solving a more complex task.

GPT-Live delegates difficult tasks

GPT-Live is designed to maintain the pace of the conversation, not to handle all the heavy work on its own. When a question requires searching the internet, deeper reasoning or completing a longer task, it delegates that work to another model.

At launch, that model will be GPT-5.5. GPT-Live will continue talking with you while it waits for the result and will be able to incorporate new generations of reasoning models as they become available.

This lets you ask for a recent fact during a conversation, request a complex comparison or get help with a multistep task without the system going silent while it works.

OpenAI says GPT-Live-1 and GPT-Live-1 mini were preferred over Advanced Voice Mode in tests involving conversations lasting between 5 and 10 minutes. The evaluations measured naturalness, turn-taking, interruptions and the overall flow of the conversation. They also recorded improvements in tests of scientific reasoning, web search and multistep phone support.

What changes in ChatGPT Voice

When users tap the voice button, they will find an updated experience with several improvements:

• More natural responses, with pauses and brief acknowledgments of what you are saying.
• A better ability to wait while you think or ask it to remain silent.
• Fewer misunderstandings caused by traffic, nearby conversations and other noise.
• A choice between the Instant, Medium and High reasoning levels.
• Visual cards with information about topics such as the weather, stocks and sports.
• Support for search, memory, images and files.

ChatGPT has also remastered its nine available voices. The system is designed for conversation, not to imitate a specific person, and uses preset voices with protections against voice impersonation.

More than 150 million people use voice and dictation features every week in ChatGPT. Uses include practicing languages, getting hands-free help, telling stories before bed and chatting during a commute.

Safety in voice conversations

OpenAI has added specific tests for voice situations involving self-harm, psychosis, mania, emotional dependency, violence and sexual content. According to the company, GPT-Live performed as well as or better than Advanced Voice Mode in almost every area evaluated.

The system can also take action while it is speaking. If it detects a potentially dangerous response, it can redirect the conversation toward a safer alternative, show support resources or end the conversation in the highest-risk cases. For self-harm situations, it includes support flows with crisis lines reviewed by experts.

Additional protections are in place for teenagers. Parents can decide whether their children use ChatGPT Voice through parental controls and, in high-risk situations involving possible self-harm, linked parents may receive a notification.

GPT-Live-1 will be the default voice model for Go, Plus and Pro users. The GPT-Live-1 mini version will be the default option for free accounts. OpenAI also plans to bring both models to its API so developers and businesses can integrate them into their own products.

For now, GPT-Live does not support video voice conversations or screen sharing in ChatGPT, although OpenAI says it is working to add those features. It also warns that performance is not the same in every language: some may have unnatural accents or lower fluency.

The importance of GPT-Live is not just that ChatGPT speaks with a more convincing voice. The change lies in combining continuous conversation with models capable of searching, reasoning and carrying out tasks in the background. What remains to be seen is whether that naturalness holds up when conversations become longer, more complex and closer to an assistant acting on your behalf.